Call intrinsics in SIMD binary operations

If the SIMD type is known to have an inner vector representation that is
known by LLVM, a fast path is used to call into a polymorphic builtin
operation.

Author: nvzqz